## 設定
verbose = False

### 言語の割合の均等化
balanced = True

### LDA 用

## トピック数
n_topics = 15 # 30は多過ぎる？

## doc, term の設定
doc_type      = 'form'
doc_attr      = 'spell'
max_doc_size  = 12
##
term_size = 'character'
term_type = 'skippy2gram'
## skippy n-gram の結合範囲
max_distance_val = round(max_doc_size * 0.8)
print(f"max_distance_val: {max_distance_val}")

## ngram を包括的にするかどうか
ngram_is_inclusive = True

### DTM 構築

## term の最低頻度
term_min_freq = 2

## 高頻度 term の濫用指標: 大きくし過ぎないように．0.05 は十分に大きい
term_abuse_threshold = 0.05

max_distance_val: 10

import sys, os, random, re, glob
import pandas as pd
import pprint as pp
from functools import reduce

## load data to process
from pathlib import Path
import pprint as pp

wd = Path(".")
##
dirs = [ x for x in wd.iterdir() if x.is_dir() and not x.match(r"plot*") ]
if verbose:
    print(f"The following {len(dirs)} directories are potential targets:")
    pp.pprint(dirs)

## list up files in target directory 
wd = Path(".")
target_dir = "data-words" # can be changed
target_files = sorted(list(wd.glob(f"{target_dir}/*.csv")))
#
print(f"\n{target_dir} contains {len(target_files)} files to process")
pp.pprint(target_files)

data-words contains 9 files to process
[PosixPath('data-words/base-sound-English-r6e-originals.csv'),
 PosixPath('data-words/base-sound-German-r1a-original.csv'),
 PosixPath('data-words/base-spell-English-r6e-originals.csv'),
 PosixPath('data-words/base-spell-Esperanto-r0-orginal.csv'),
 PosixPath('data-words/base-spell-French-r0-originals.csv'),
 PosixPath('data-words/base-spell-German-r1a-originals.csv'),
 PosixPath('data-words/base-spell-Icelandic-r0-original.csv'),
 PosixPath('data-words/base-spell-Russian-r0-originals.csv'),
 PosixPath('data-words/base-spell-Swahili-r0-orginal.csv')]

import pandas as pd

## データ型の辞書
types = "spell sound freq".split(" ")
type_setting = { t : 0 for t in types }
print(type_setting)

## 言語名の辞書
langs = "english esperanto french german icelandic russian swahili".split(" ")
#langs = "english esperanto french german russian swahili".split(" ")
#langs = "english esperanto french german icelandic swahili".split(" ")
lang_setting = { lang : 0 for lang in langs }
print(lang_setting)

## 辞書と統合
settings = { 'form': None, **type_setting, **lang_setting }
print(settings)

{'spell': 0, 'sound': 0, 'freq': 0}
{'english': 0, 'esperanto': 0, 'french': 0, 'german': 0, 'icelandic': 0, 'russian': 0, 'swahili': 0}
{'form': None, 'spell': 0, 'sound': 0, 'freq': 0, 'english': 0, 'esperanto': 0, 'french': 0, 'german': 0, 'icelandic': 0, 'russian': 0, 'swahili': 0}

vars = list(settings.keys())
print(f"targe var names: {vars}")

d_parts = [ ]
for lang in langs:
    local_settings = settings.copy()
    print(f"processing: {lang}")
    try:
        for f in [ f for f in target_files if lang.capitalize() in str(f) ]:
            print(f"reading: {f}")
            # 言語名の指定
            local_settings[lang] = 1
            # 型名の指定
            for type in vars:
                if type in str(f):
                    local_settings[type] = 1
            #
            d = pd.read_csv(f, encoding='utf-8', sep = ",", on_bad_lines = 'skip') # Crucially, ...= skip
            df = pd.DataFrame(d, columns = vars)
            for var in [ var for var in (types + langs) if var != 'freq' ]:
                df[var] = local_settings[var]
            d_parts.append(df)
    except IndexError:
        pass
#
if verbose:
    d_parts

targe var names: ['form', 'spell', 'sound', 'freq', 'english', 'esperanto', 'french', 'german', 'icelandic', 'russian', 'swahili']
processing: english
reading: data-words/base-sound-English-r6e-originals.csv
reading: data-words/base-spell-English-r6e-originals.csv
processing: esperanto
reading: data-words/base-spell-Esperanto-r0-orginal.csv
processing: french
reading: data-words/base-spell-French-r0-originals.csv
processing: german
reading: data-words/base-sound-German-r1a-original.csv
reading: data-words/base-spell-German-r1a-originals.csv
processing: icelandic
reading: data-words/base-spell-Icelandic-r0-original.csv
processing: russian
reading: data-words/base-spell-Russian-r0-originals.csv
processing: swahili
reading: data-words/base-spell-Swahili-r0-orginal.csv

## データ統合

raw_df = pd.concat(d_parts)
raw_df

## 文字数の列を追加

raw_df['size'] = [ len(x) for x in raw_df[doc_type] ]
raw_df

## 言語名= language の列を追加

check = False
language_vals = [ ]
for i, row in raw_df.iterrows():
    if check:
        print(row)
    for j, lang in enumerate(langs):
        if check:
            print(f"{i}: {lang}")
        if row[lang] == 1:
            language_vals.append(lang)
if verbose:
    print(language_vals)
    len(language_vals)
#
raw_df['language'] = language_vals
raw_df

## 言語の選別

select_languages = True

selected_langs = re.split(r",\s*", "english, french, german, russian, swahili")
print(f"selected languages: {selected_langs}")

if select_languages:
    df_new = [ ]
    for lang in selected_langs:
        df_new.append(raw_df[raw_df[lang] == 1])
    raw_df = pd.concat(df_new)
#
raw_df

selected languages: ['english', 'french', 'german', 'russian', 'swahili']

## 文字数の分布
import numpy as np
import matplotlib.pyplot as plt

fig = plt.figure()
ax = fig.add_subplot(1,1,1)

ax.hist(raw_df['size'], bins = 40)
ax.set_xlabel('length of doc')
ax.set_ylabel('freq')
plt.title(f"Length distribution for docs")
fig.show()

/var/folders/s2/lk8hdt6j10j0xyycw1lbjsm40000gn/T/ipykernel_55916/1088473461.py:12: UserWarning: FigureCanvasAgg is non-interactive, and thus cannot be shown
  fig.show()

## 長さで濾過

print(f"max doc size: {max_doc_size}")
original_size = len(raw_df)
raw_df = raw_df[raw_df['size'] < max_doc_size]
filtered_size = len(raw_df)
print(f"{original_size - filtered_size} cases removed")

max doc size: 12
296 cases removed

## 結果の検査 1

for lang in langs:
    print(raw_df[lang].value_counts())

english
1    8249
0    4228
Name: count, dtype: int64
esperanto
0    12477
Name: count, dtype: int64
french
0    11492
1      985
Name: count, dtype: int64
german
0    10912
1     1565
Name: count, dtype: int64
icelandic
0    12477
Name: count, dtype: int64
russian
0    11504
1      973
Name: count, dtype: int64
swahili
0    11772
1      705
Name: count, dtype: int64

## 結果の検査 2

for type in types:
    print(raw_df[type].value_counts())

spell
1    7588
0    4889
Name: count, dtype: int64
sound
1    9814
0    2663
Name: count, dtype: int64
freq
1                  11495
1                    966
1 не                   1
1 то время как         1
1 северу               1
1 него                 1
1 будет                1
1 образом              1
1 мышь                 1
Name: count, dtype: int64

## 統合: 割合補正を適用

eng_reduct_factor = 0.2
if balanced:
    eng_df = raw_df[raw_df['english'] == 1]
    non_eng_df = raw_df[raw_df['english'] == 0]
    eng_reduced_df = eng_df.sample(round(len(eng_df) * eng_reduct_factor))
    raw_df = pd.concat([eng_reduced_df, non_eng_df])
raw_df

## 結果の検査 3

for lang in langs:
    print(raw_df[lang].value_counts())

english
0    4228
1    1650
Name: count, dtype: int64
esperanto
0    5878
Name: count, dtype: int64
french
0    4893
1     985
Name: count, dtype: int64
german
0    4313
1    1565
Name: count, dtype: int64
icelandic
0    5878
Name: count, dtype: int64
russian
0    4905
1     973
Name: count, dtype: int64
swahili
0    5173
1     705
Name: count, dtype: int64

## 順序のランダマイズ

import sklearn.utils
raw_df = sklearn.utils.shuffle(raw_df)

## データ名の指定
df = raw_df[raw_df[doc_attr] == 1]
print(f"doc_attr: {doc_attr}")
df

doc_attr: spell

df[df['english'] == 1]

## ngram の追加

import sys
sys.path.append('..')

import re
import ngrams
import importlib
importlib.reload(ngrams)
import ngrams_skippy

bases = df[doc_type]

## 1gram 列の追加
#sep = r""
#unigrams = [ list(filter(lambda x: len(x) > 0, y)) for y in [ re.split(sep, z) for z in bases ] ]
unigrams = ngrams.gen_unigrams(bases, sep = r"", check = False)
if verbose:
    random.sample(unigrams, 5)
#
df['1gram'] = unigrams
#df.loc[:,'1gram'] = unigrams
df

/var/folders/s2/lk8hdt6j10j0xyycw1lbjsm40000gn/T/ipykernel_55916/1248262955.py:21: SettingWithCopyWarning: 
A value is trying to be set on a copy of a slice from a DataFrame.
Try using .loc[row_indexer,col_indexer] = value instead

See the caveats in the documentation: https://pandas.pydata.org/pandas-docs/stable/user_guide/indexing.html#returning-a-view-versus-a-copy
  df['1gram'] = unigrams

## 2gram列の追加

bigrams = ngrams.gen_bigrams(bases, sep = r"", check = False)

## 包括的 2gram の作成
if ngram_is_inclusive:
    bigrams = [ [*b, *u] for b, u in zip(bigrams, unigrams) ]

if verbose:
    print(random.sample(bigrams, 3))

df['2gram'] = bigrams
if verbose:
    df

/var/folders/s2/lk8hdt6j10j0xyycw1lbjsm40000gn/T/ipykernel_55916/1480305306.py:1: SettingWithCopyWarning: 
A value is trying to be set on a copy of a slice from a DataFrame.
Try using .loc[row_indexer,col_indexer] = value instead

See the caveats in the documentation: https://pandas.pydata.org/pandas-docs/stable/user_guide/indexing.html#returning-a-view-versus-a-copy
  df['2gram'] = bigrams

## 3gram列の追加

trigrams = ngrams.gen_trigrams(bases, sep = r"", check = False)

## 包括的 3gram の作成
if ngram_is_inclusive:
    trigrams = [ [ *t, *b ] for t, b in  zip(trigrams, bigrams) ]

if verbose:
    print(random.sample(trigrams, 3))

df['3gram'] = trigrams
if verbose:
    df

/var/folders/s2/lk8hdt6j10j0xyycw1lbjsm40000gn/T/ipykernel_55916/3715201492.py:1: SettingWithCopyWarning: 
A value is trying to be set on a copy of a slice from a DataFrame.
Try using .loc[row_indexer,col_indexer] = value instead

See the caveats in the documentation: https://pandas.pydata.org/pandas-docs/stable/user_guide/indexing.html#returning-a-view-versus-a-copy
  df['3gram'] = trigrams

## skippy 2grams の生成

import sys
sys.path.append("..") # library path に一つ上の階層を追加

import ngrams_skippy

skippy_2grams = [ ngrams_skippy.generate_skippy_bigrams(x,
                                                         missing_mark = '…',
                                                         max_distance = max_distance_val, check = False)
                  for x in df['1gram'] ]

## 包括的 skippy 2-grams の生成
if ngram_is_inclusive:
    for i, b2 in enumerate(skippy_2grams):
        b2.extend(unigrams[i])
#
if verbose:
    random.sample(skippy_2grams, 3)

## skippy 2gram 列の追加

df['skippy2gram'] = skippy_2grams
df

/var/folders/s2/lk8hdt6j10j0xyycw1lbjsm40000gn/T/ipykernel_55916/3263801935.py:3: SettingWithCopyWarning: 
A value is trying to be set on a copy of a slice from a DataFrame.
Try using .loc[row_indexer,col_indexer] = value instead

See the caveats in the documentation: https://pandas.pydata.org/pandas-docs/stable/user_guide/indexing.html#returning-a-view-versus-a-copy
  df['skippy2gram'] = skippy_2grams

## skippy 3grams の生成

import sys
sys.path.append("..") # library path に一つ上の階層を追加

import ngrams_skippy

skippy_3grams = [ ngrams_skippy.generate_skippy_trigrams(x,
                                                         missing_mark = '…',
                                                         max_distance = max_distance_val, check = False)
                  for x in df['1gram'] ]

## 包括的 skippy 3-grams の生成
if ngram_is_inclusive:
    for i, t2 in enumerate(skippy_3grams):
        t2.extend(skippy_2grams[i])
#
if verbose:
    random.sample(skippy_3grams, 3)

## skippy 3gram 列の追加

df['skippy3gram'] = skippy_3grams
df

/var/folders/s2/lk8hdt6j10j0xyycw1lbjsm40000gn/T/ipykernel_55916/1159231133.py:3: SettingWithCopyWarning: 
A value is trying to be set on a copy of a slice from a DataFrame.
Try using .loc[row_indexer,col_indexer] = value instead

See the caveats in the documentation: https://pandas.pydata.org/pandas-docs/stable/user_guide/indexing.html#returning-a-view-versus-a-copy
  df['skippy3gram'] = skippy_3grams

## LDA 構築の基になる document-term matrix (dtm) を構築

from gensim.corpora.dictionary import Dictionary

bots = df[term_type]
diction = Dictionary(bots)

## 結果の確認
print(diction)

Dictionary<2935 unique tokens: ['i', 'is', 'i…t', 'i…é', 'r']...>

## diction の濾過

import copy
diction_copy = copy.deepcopy(diction)

## filter適用: 実は諸刃の刃で，token数が少ない時には適用しない方が良い
print(f"min freq filter: {term_min_freq}")
print(f"abuse filter: {term_abuse_threshold}")

apply_filter = True
if apply_filter:
    diction_copy.filter_extremes(no_below = term_min_freq, no_above = term_abuse_threshold)

## check
print(diction_copy)

min freq filter: 2
abuse filter: 0.05
Dictionary<2258 unique tokens: ['is', 'i…t', 'i…é', 'ré', 'r…i']...>

## Corpus (gensim の用語では corpus) の構築

corpus = [ diction.doc2bow(bot) for bot in bots ]

## check
check = True
if verbose:
    sample_n = 5
    print(random.sample(corpus, sample_n))
#
print(f"Number of documents: {len(corpus)}")

Number of documents: 4256

## LDA モデルの構築

from gensim.models import LdaModel
#from tqdm import tqdm

## LDAモデル
print(f"Building LDA model with n_topics: {n_topics}")
lda = LdaModel(corpus, id2word = diction, num_topics = n_topics, alpha = 0.01)
#
print(lda) # print(..)しないと中身が見れない

Building LDA model with n_topics: 15
LdaModel<num_terms=2935, num_topics=15, decay=0.5, chunksize=2000>

%%capture --no-display

## LDA のtopic ごとに，関連度の高い term を表示

import pandas as pd

n_terms = 20 # topic ごとに表示する term 数の指定
topic_dfs = [ ]
for topic in range(n_topics):
    terms = [ ]
    for i, prob in lda.get_topic_terms(topic, topn = n_terms):
        terms.append(diction.id2token[ int(i) ])
    #
    topic_dfs.append(pd.DataFrame([terms], index = [ f'topic {topic+1}' ]))
#
topic_term_df = pd.concat(topic_dfs)

## Table で表示
topic_term_df.T

%%capture --no-display

## pyLDAvis を使った結果 LDA の可視化: 階層クラスタリングより詳しい

import pyLDAvis

#installed_version = sys.version
installed_version = pyLDAvis.__version__

print(f"installed_version: {installed_version}")

if float(installed_version[:3]) > 3.1:
    import pyLDAvis.gensim_models as gensimvis
else:
    import pyLDAvis.gensim as gensimvis
#
pyLDAvis.enable_notebook()

#
lda_used     = lda
corpus_used  = corpus
diction_used = diction

## 実行パラメター
use_tSNE = False
if use_tSNE:
    vis = gensimvis.prepare(lda_used, corpus_used, diction_used, mds = 'tsne',
                            n_jobs = 1, sort_topics = False)
else:
    vis = gensimvis.prepare(lda_used, corpus_used, diction_used,
                            n_jobs = 1, sort_topics = False)
#
pyLDAvis.display(vis)

## 結果について
## topic を表わす円の重なりが多いならn_topics が多過ぎる可能性がある．
## ただし2Dで重なっていても，3Dなら重なっていない可能性もある

## LDA がD に対して生成した topics の弁別性を確認

## 得られたtopics を確認
topic_dist = lda.get_topics()
if verbose:
    topic_dist

## 検査 1: topic ごとに分布の和を取る

print(topic_dist.sum(axis = 1))

[0.99999994 0.99999994 1.0000001  1.         1.0000001  1.
 1.0000001  0.99999994 0.99999994 1.0000001  1.         1.
 1.         1.         0.9999998 ]

## 検査 2: 総和を求める: n_topics にほぼ等しいなら正常

print(topic_dist.sum())

15.0

## term エンコード値の分布を確認

import matplotlib.pyplot as plt

plt.figure(figsize = (4,5))

sampling_rate = 0.3
df_size = len(topic_dist)
sample_n = round(df_size * sampling_rate)
topic_sampled = random.sample(list(topic_dist), sample_n)

T = sorted([ sorted(x, reverse = True) for x in topic_sampled ])

plt.plot(T, range(len(T)))
plt.title("Distribution of sorted values ({sample_n} samples) for topic/term encoding")
plt.show()

## tSNE を使った topics のグループ化 (3D)

from sklearn.manifold import TSNE
import numpy as np

## tSNE のパラメターを設定
## n_components は射影先の空間の次元: n_components = 3 なら3次元空間に射影
## perplexity は結合の強さを表わす指数で，値に拠って結果が代わるので，色々な値を試すと良い

#perplexity_val = 10 # 大き過ぎると良くない
top_perplexity_reduct_rate = 0.3
perplexity_val = round(len(topic_dist) * top_perplexity_reduct_rate)
topic_tSNE_3d = TSNE(n_components = 3, random_state = 0, perplexity = perplexity_val, n_iter = 1000)

## データに適用
top_tSNE_3d_fitted = topic_tSNE_3d.fit_transform(np.array(topic_dist))

## Plotlyを使って tSNE の結果の可視化 (3D)

#import plotly.express as pex
import plotly.graph_objects as go
import numpy as np

top_tSNE = top_tSNE_3d_fitted
fig = go.Figure(data = [go.Scatter3d(x = top_tSNE[:,0], y = top_tSNE[:,1], z = top_tSNE[:,2],
                                     mode = 'markers')])

## 3D 散布図にラベルを追加する処理は未実装
title_val = f"3D tSNE view for LDA (#topics: {n_topics}, doc: {doc_type}, term: {term_size} {term_type})"
fig.update_layout(autosize = False,
                  width = 600, height = 600, title = title_val)
fig.show()

## 構築した LDA モデルを使って文(書)を分類する

## .get_document_topics(..) は　minimu_probability = 0としないと
## topic の値が小さい場合に値を返さないので，

## パラメター

ntopics = n_topics # LDA の構築の最に指定した値を使う

check      = False
encoding = [ ]
for i, row in df.iterrows():
    if check:
        print(f"row: {row}")
    doc = row[doc_type]
    bot = row[term_type]
    ## get_document_topics(..) では　minimu_probability = 0 としないと
    ## 値が十分に大きな topics に関してだけ値が取れる
    enc = lda.get_document_topics(diction.doc2bow(bot), minimum_probability = 0)
    if check:
        print(f"enc: {enc}")
    encoding.append(enc)
#
len(encoding)

4256

## enc 列の追加

#df['enc'] = np.array(encoding) # This flattens arrays
#df['enc'] = list(encoding) # ineffective
df['enc'] = [ list(map(lambda x: x[1], y)) for y in encoding ]
if verbose:
    df['enc']

/var/folders/s2/lk8hdt6j10j0xyycw1lbjsm40000gn/T/ipykernel_55916/1047258704.py:5: SettingWithCopyWarning:


A value is trying to be set on a copy of a slice from a DataFrame.
Try using .loc[row_indexer,col_indexer] = value instead

See the caveats in the documentation: https://pandas.pydata.org/pandas-docs/stable/user_guide/indexing.html#returning-a-view-versus-a-copy

## エンコーディングのstd の分布を見る

from scipy.stats import tstd
from matplotlib import pyplot as plt

plt.figure(figsize = (6,4))

std_data = [ tstd(x) for x in df['enc'] ]
plt.hist(std_data)

plt.title("Distribution of standard deviations")
plt.show()

## doc のエンコーディング

## 一様分布の事例を除外

from scipy.stats import tstd # standard deviation の計算用

print(f"{len(df)} instances before filtering")

check = False

doc_enc = df['enc']
max_std = max([ tstd(x) for x in doc_enc])
if check: print(f"std max: {max_std}")

min_std = min([ tstd(x) for x in doc_enc])
if check: print(f"std min: {min_std}")

first_min_std = list(sorted(set([ tstd(x) for x in doc_enc])))[-0]
print(f"std 1st min: {first_min_std}")

second_min_std = list(sorted(set([ tstd(x) for x in doc_enc])))[-1]
print(f"std 2nd min: {second_min_std}")

4256 instances before filtering
std 1st min: 0.08733703030759321
std 2nd min: 0.2576133771310708

## df_filtered の定義

## 閾値は2番目に小さい値より小さく最小値よりは大きな値であるべき
std_threshold = second_min_std / 4 # 穏健な値を得るために4で割った
print(f"std_threshold: {std_threshold}")

## Rっぽい次のコードは通らない
#df_filtered = df[ df['encoding'] > std_threshold ]

## 通るのは次のコード: Creating a list of True/False and apply it to DataFrame 
std_tested = [ False if tstd(x) < std_threshold else True for x in df['enc'] ]
df_filtered = df[ std_tested ]
#
print(f"{len(df_filtered)} instances after filtering ({len(df) - len(df_filtered)} instances removed)")

std_threshold: 0.0644033442827677
4256 instances after filtering (0 instances removed)

## doc エンコード値の分布を確認

sample_n = 50
E = sorted([ sorted(x, reverse = True) for x in df_filtered['enc'].sample(sample_n) ])

plt.figure(figsize = (5,5))
plt.plot(E, range(len(E)))
plt.title(f"Distribution of sorted encoding values for sampled {sample_n} docs")
plt.show()

len(df_filtered['language'])

4256

## tSNE 用の事例サンプリング = tSNE_df の定義

tSNE_sampling = True
tSNE_sampling_rate = 0.33
if tSNE_sampling:
    tSNE_df_original = df_filtered.copy()
    sample_n = round(len(tSNE_df_original) * tSNE_sampling_rate)
    tSNE_df = tSNE_df_original.sample(sample_n)
    print(f"tSNE_df has {len(tSNE_df)} rows after sampling")
else:
    tSNE_df = df_filtered

tSNE_df has 1404 rows after sampling

tSNE_df.columns

Index(['form', 'spell', 'sound', 'freq', 'english', 'esperanto', 'french',
       'german', 'icelandic', 'russian', 'swahili', 'size', 'language',
       '1gram', '2gram', '3gram', 'skippy2gram', 'skippy3gram', 'enc'],
      dtype='object')

## tSNE の結果の可視化: Plotly を使った 3D 描画

import numpy as np
from sklearn.manifold import TSNE as tSNE

import plotly.express as pex
import plotly.graph_objects as go
import matplotlib.pyplot as plt

## tSNE のパラメターを設定

perplexity_max_val = round(len(tSNE_df)/4)

for perplexity_val in range(5, perplexity_max_val, 30):
    
    ## tSNE 事例の生成
    tSNE_3d_varied = tSNE(n_components = 3, random_state = 0, perplexity = perplexity_val, n_iter = 1000)

    ## データに適用
    doc_enc = np.array(list(tSNE_df['enc']))
    doc_tSNE_3d_varied = tSNE_3d_varied.fit_transform(doc_enc)

    T = zip(doc_tSNE_3d_varied[:,0], doc_tSNE_3d_varied[:,1], doc_tSNE_3d_varied[:,2],
            tSNE_df['language']) # zip(..)が必要
    df = pd.DataFrame(T, columns = ['D1', 'D2', 'D3', 'language'])

    ## 作図
    fig = go.Figure()
    for lang in np.unique(df['language']):
        part = df[df['language'] == lang]
        fig.add_trace(
            go.Scatter3d(
                x = part['D1'], y = part['D2'], z = part['D3'],
                name = lang, mode = 'markers', marker = dict(size = 6),
                showlegend = True
            )
        )

    title_val = f"tSNE 3D map (ppl: {perplexity_val}) of {doc_attr}s encoded by LDA ({n_topics} topics; term: {term_type})"
    fig.update_layout(title = dict(text = title_val),
                      autosize = False, width = 600, height = 600,)
    fig.show()

## 階層クラスタリングのための事例のサンプリング

hc_sampling_rate = 0.1 # 大きくし過ぎると図が見にくい
df_size = len(tSNE_df)
hc_sample_n = round(df_size * hc_sampling_rate)
hc_df = tSNE_df.sample(hc_sample_n)
##
print(f"{hc_sample_n} rows are sampled")
hc_df['language'].value_counts()

140 rows are sampled

language
french     38
russian    29
german     26
swahili    25
english    22
Name: count, dtype: int64

## doc 階層クラスタリングの実行

import numpy as np
import plotly
import matplotlib.pyplot as plt

from scipy.cluster.hierarchy import dendrogram, linkage

## 距離行列の生成
Enc = list(hc_df['enc'])
linkage = linkage(Enc, method = 'ward', metric = 'euclidean')

## 描画サイズの指定
plt.figure(figsize = (5, round(len(hc_df) * 0.15))) # This needs to be run here, before dendrogram construction.

## 事例ラベルの生成
label_vals = [ x[:max_doc_size] for x in list(hc_df[doc_type]) ] # truncate doc keys

## 樹状分岐図の作成
dendrogram(linkage, orientation = 'left', labels = label_vals, leaf_font_size = 7)

## 描画
plt.title(f"Hierarchical clustering of (sampled) {len(hc_df)} (= {100 * hc_sampling_rate}%) {doc_attr}s as docs\n \
    encoded via LDA ({n_topics} topics) with {term_type} as terms")

## ラベルに language に対応する色を付ける
lang_colors = { lang_name : i for i, lang_name in enumerate(np.unique(hc_df['language'])) }
ax = plt.gca()
for ticker in ax.get_ymajorticklabels():
    form = ticker.get_text()
    row = hc_df.loc[hc_df[doc_type] == form]
    #lang = row['language']
    lang = row['language'].to_string().split()[-1] # trick
    try:
        lang_id = lang_colors[lang]
    except (TypeError, KeyError):
        print(f"color encoding error at: {lang}")
    #
    ticker.set_color(plotly.colors.qualitative.Plotly[lang_id]) # id の基数調整
#
plt.show()

## tSNE の結果の可視化 (2D)

#import seaborn as sns
import numpy as np
import plotly
import plotly.express as pex
import matplotlib.pyplot as plt
from adjustText import adjust_text

## tSNE 事例の生成
perplexity_selected = 250
tSNE_3d = tSNE(n_components = 3, random_state = 0, perplexity = perplexity_selected, n_iter = 1000)

## データに適用
doc_enc = np.array(list(tSNE_df['enc']))
doc_tSNE_3d = tSNE_3d.fit_transform(doc_enc)

T = zip(doc_tSNE_3d[:,0], doc_tSNE_3d[:,1], doc_tSNE_3d[:,2],
        tSNE_df['language']) # zip(..)が必要
df = pd.DataFrame(T, columns = ['D1', 'D2', 'D3', 'language'])

## 描画
plt.figure(figsize = (5, 5))
plt.set_colors = pex.colors.qualitative.Plotly
    
for r in [ np.roll([0,1,2], -i) for i in range(0,3) ]:
    if check:
        print(r)
    X, Y = df.iloc[:,r[0]], df.iloc[:,r[1]]
    gmax = max(X.max(), Y.max())
    gmin = min(X.min(), Y.min())
    plt.xlim(gmin, gmax)
    plt.ylim(gmin, gmax)

    colormap = pex.colors.qualitative.Plotly
    lang_list = list(np.unique(tSNE_df['language']))
    cmapped = [ colormap[lang_list.index(lang)] for lang in df['language'] ]
    scatter = plt.scatter(X, Y, s = 40, c = cmapped, edgecolors = 'w')

    ## 文字を表示する事例のサンプリング
    lab_sampling_rate = 0.02
    lab_sample_n = round(len(tSNE_df) * lab_sampling_rate)
    sampled_keys = [ doc[:max_doc_size] for doc in random.sample(list(tSNE_df[doc_type]), lab_sample_n) ]

    ## labels の生成
    texts = [ ]
    for x, y, s in zip(X, Y, sampled_keys):
        texts.append(plt.text(x, y, s, size = 9, color = 'blue'))

    ## label に repel を追加: adjustText package の導入が必要
    adjust_text(texts, force_points = 0.2, force_text = 0.2,
                expand_points = (1, 1), expand_text = (1, 1),
                arrowprops = dict(arrowstyle = "-", color = 'black', lw = 0.5))
    #

    plt.title(f"tSNE (ppl: {perplexity_selected}) 2D map of {len(tSNE_df)} {doc_attr}s via LDA ({term_type}; {n_topics} topics)")
    #plt.legend(np.unique(cmapped))
    plt.show()

	form	spell	sound	freq	english	esperanto	french	german	icelandic	russian	swahili	size	language	1gram	2gram	3gram	skippy2gram	skippy3gram
427	résisté	1	0	1.0	0	0	1	0	0	0	0	7	french	[r, é, s, i, s, t, é]	[ré, és, si, is, st, té, r, é, s, i, s, t, é]	[rés, ési, sis, ist, sté, ré, és, si, is, st, ...	[ré, r…s, r…i, r…t, r…é, és, é…i, é…s, é…t, é…...	[rés, ré…i, ré…s, ré…t, ré…é, r…si, r…s…s, r…s...
495	na	1	0	1	0	0	0	0	0	0	1	2	swahili	[n, a]	[na, n, a]	[n, a, na, n, a]	[na, n, a]	[na, na, n, a]
533	nyeusi	1	0	1	0	0	0	0	0	0	1	6	swahili	[n, y, e, u, s, i]	[ny, ye, eu, us, si, n, y, e, u, s, i]	[nye, yeu, eus, usi, ny, ye, eu, us, si, n, y,...	[ny, n…e, n…u, n…s, n…i, ye, y…u, y…s, y…i, eu...	[nye, ny…u, ny…s, ny…i, n…eu, n…e…s, n…e…i, n…...
705	racine	1	0	1.0	0	0	1	0	0	0	0	6	french	[r, a, c, i, n, e]	[ra, ac, ci, in, ne, r, a, c, i, n, e]	[rac, aci, cin, ine, ra, ac, ci, in, ne, r, a,...	[ra, r…c, r…i, r…n, r…e, ac, a…i, a…n, a…e, ci...	[rac, ra…i, ra…n, ra…e, r…ci, r…c…n, r…c…e, r…...
768	récolte	1	0	1.0	0	0	1	0	0	0	0	7	french	[r, é, c, o, l, t, e]	[ré, éc, co, ol, lt, te, r, é, c, o, l, t, e]	[réc, éco, col, olt, lte, ré, éc, co, ol, lt, ...	[ré, r…c, r…o, r…l, r…t, r…e, éc, é…o, é…l, é…...	[réc, ré…o, ré…l, ré…t, ré…e, r…co, r…c…l, r…c...
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
130	phrase	1	0	1.0	0	0	1	0	0	0	0	6	french	[p, h, r, a, s, e]	[ph, hr, ra, as, se, p, h, r, a, s, e]	[phr, hra, ras, ase, ph, hr, ra, as, se, p, h,...	[ph, p…r, p…a, p…s, p…e, hr, h…a, h…s, h…e, ra...	[phr, ph…a, ph…s, ph…e, p…ra, p…r…s, p…r…e, p…...
257	licht	1	1	1	0	0	0	1	0	0	0	5	german	[l, i, c, h, t]	[li, ic, ch, ht, l, i, c, h, t]	[lic, ich, cht, li, ic, ch, ht, l, i, c, h, t]	[li, l…c, l…h, l…t, ic, i…h, i…t, ch, c…t, ht,...	[lic, li…h, li…t, l…ch, l…c…t, l…ht, ich, ic…t...
165	vor	1	1	1	0	0	0	1	0	0	0	3	german	[v, o, r]	[vo, or, v, o, r]	[vor, vo, or, v, o, r]	[vo, v…r, or, v, o, r]	[vor, vo, v…r, or, v, o, r]
562	beginn	1	1	1	0	0	0	1	0	0	0	6	german	[b, e, g, i, n, n]	[be, eg, gi, in, nn, b, e, g, i, n, n]	[beg, egi, gin, inn, be, eg, gi, in, nn, b, e,...	[be, b…g, b…i, b…n, eg, e…i, e…n, gi, g…n, in,...	[beg, be…i, be…n, b…gi, b…g…n, b…in, b…i…n, b…...
133	dire	1	0	1.0	0	0	1	0	0	0	0	4	french	[d, i, r, e]	[di, ir, re, d, i, r, e]	[dir, ire, di, ir, re, d, i, r, e]	[di, d…r, d…e, ir, i…e, re, d, i, r, e]	[dir, di…e, d…re, ire, di, d…r, d…e, ir, i…e, ...

	topic 1	topic 2	topic 3	topic 4	topic 5	topic 6	topic 7	topic 8	topic 9	topic 10	topic 11	topic 12	topic 13	topic 14	topic 15
0	a	в	i	a	s	ый	о	e	е	e	з	а	о	u	c
1	e	u	r	u	e	й	л	s	о	n	з…а	с	д	l	u
2	t	о	n	e	o	ы	к	c	р	o	а	и	а	i	c…u
3	r	у…в	a	m	t	ны	е	n	а	i	ь	л	х…т	a	a
4	p	e	e	r	r	и…с	а	i	т	t	и	т	р	k	t
5	h	ов	i…e	n	a	н…й	i	e…e	и	en	r	о	и	e	e
6	s	r	m	a…e	n	т	j	h	в	e…e	e	с…а	г	k…a	c…r
7	o	во	i…i	l	f	с	д	l	п	e…n	зн	н	х	ku	h
8	a…e	тв	o	d	u	о…ы	l	en	н	o…n	н	р	r	u…a	co
9	b	t	i…r	a…a	i	я	д…л	s…e	ь	n…e	и…ь	я	п…х	s	l
10	t…e	у…о	s	g	h	о…с	б	ch	к	i…n	ть	ал	га	o	ch
11	a…t	p	d	p	c	и…я	н	c…e	л	g	з…ь	ст	од	k…i	c…e
12	n	вс	in	ma	s…e	х	e	i…e	е…ь	o…t	за	б	т	le	c…a
13	t…r	л	is	i	s…n	п	ло	r	о…е	b	з…т	и…а	м…й	m	r
14	h…r	ou	i…a	m…a	st	o	о…о	n…e	р…т	t…n	жи	с…л	о…а	n	c…t
15	a…r	u…e	a…i	a…r	o…s	н…ы	л…а	s…n	п…о	d	or	ра	х…и	u…i	é
16	é	о…ь	i…o	f	s…i	л	a	h…e	о…т	s	ж…ь	ы	р…о	il	c…l
17	l	с…в	k	t	s…u	н	с	e…n	м	l	зи	к	н	i…a	ha
18	p…r	ue	ai	al	o…e	о…й	m	h…n	п…т	on	po	а…т	wo	ll	h…u
19	er	p…u	ir	a…d	t…e	р…й	м	t	пр	o…e	т	о…т	п…д	l…e	o

	form	spell	sound	freq	english	esperanto	french	german	icelandic	russian	swahili
0	æbəhəlimə	0	1	1	1	0	0	0	0	0	0
1	ædmaɪə	0	1	1	1	0	0	0	0	0	0
2	ædmɪʃən	0	1	1	1	0	0	0	0	0	0
3	ædvæntɪdʒ	0	1	1	1	0	0	0	0	0	0
4	ædvaɪs	0	1	1	1	0	0	0	0	0	0
...	...	...	...	...	...	...	...	...	...	...	...
703	zaidi	1	0	1	0	0	0	0	0	0	1
704	ziara	1	0	1	0	0	0	0	0	0	1
705	zima	1	0	1	0	0	0	0	0	0	1
706	ziwa	1	0	1	0	0	0	0	0	0	1
707	zoezi	1	0	1	0	0	0	0	0	0	1

	form	spell	sound	freq	english	esperanto	french	german	icelandic	russian	swahili	size	language
3215	ədɪʃən	0	1	1	1	0	0	0	0	0	0	6	english
710	faɪəɹ	0	1	1	1	0	0	0	0	0	0	5	english
1222	endless	1	1	1.0	1	0	0	0	0	0	0	7	english
2511	stɹoʊk	0	1	1	1	0	0	0	0	0	0	6	english
1376	father	1	1	1.0	1	0	0	0	0	0	0	6	english
...	...	...	...	...	...	...	...	...	...	...	...	...	...
703	zaidi	1	0	1	0	0	0	0	0	0	1	5	swahili
704	ziara	1	0	1	0	0	0	0	0	0	1	5	swahili
705	zima	1	0	1	0	0	0	0	0	0	1	4	swahili
706	ziwa	1	0	1	0	0	0	0	0	0	1	4	swahili
707	zoezi	1	0	1	0	0	0	0	0	0	1	5	swahili

	form	spell	sound	freq	english	esperanto	french	german	icelandic	russian	swahili	size	language
671	coast	1	1	1.0	1	0	0	0	0	0	0	5	english
1557	gabbing	1	1	1.0	1	0	0	0	0	0	0	7	english
644	classify	1	1	1.0	1	0	0	0	0	0	0	8	english
1523	freia	1	1	1.0	1	0	0	0	0	0	0	5	english
1024	didn't	1	1	1.0	1	0	0	0	0	0	0	6	english
...	...	...	...	...	...	...	...	...	...	...	...	...	...
4233	worship	1	1	1.0	1	0	0	0	0	0	0	7	english
4067	visitor	1	1	1.0	1	0	0	0	0	0	0	7	english
297	bathe	1	1	1.0	1	0	0	0	0	0	0	5	english
1710	handle	1	1	1.0	1	0	0	0	0	0	0	6	english
1836	hourly	1	1	1.0	1	0	0	0	0	0	0	6	english